PyTorch analytical chip placer — macro-aware density modeling with two-phase overlap optimization (λ_overlap 0.1 → 15.0).
8.82% wirelength cut · 99.14% overlap-area eliminated vs naive radial init on synthetic toy (2 macros + 30 std cells, seed 1003) · 6/6 tests · CI: green
docker compose up --build → writes before/after plots + metrics into assets/.
| BEFORE (naive radial init) | AFTER (GradPlace, 2000 epochs, CPU) |
|---|---|
 |
 |
Red cells = involved in ≥1 pairwise overlap. Purple = macros. Blue = standard cells.
Repo: github.com/ArchanaChetan07/GradPlace
GradPlace is a differentiable VLSI placement loop in PyTorch: generate a random macro/std-cell netlist, scatter cells (naive baseline), then Adam-optimize (x, y) to trade off smooth Manhattan wirelength against density/overlap repulsion.
The default demo benchmark is a synthetic toy netlist (2 macros + 30 standard cells, seed 1003) — not an ISPD/contest industry case. Framing it as such keeps the metrics honest.
Entry points:
python scripts/eval_before_after.py— before/after metrics + plots (source of README numbers)python test.py --quick— multi-size synthetic suitepython main.py --quick— thin wrapper (CPU by default)
| Knob | Value | What it does |
|---|---|---|
macro_weight |
5.0 | Macros count 5× in the density map so big blocks don’t get buried by tiny cells |
| Std-cell bin cap | 0.25 × bin_capacity | Limits how much density a single std-cell can dump into one bin |
λ_overlap ramp |
0.1 → 1.0 → 15.0 | Phase 1 (first 40% of epochs) slowly turns up overlap pressure; phase 2 finishes at 15 |
| Wirelength | Smooth Manhattan (α=0.1) |
Differentiable routing-cost proxy between connected pins |
| Optimizer | Adam, lr=0.01 |
Updates cell centers only |
flowchart LR
GEN["Synthetic netlist"] --> INIT["Radial naive init"]
INIT --> OPT["Adam · WL + λ_ov·overlap"]
OPT --> OUT["Coords + metrics + plots"]
Command: python scripts/eval_before_after.py --epochs 2000 --device cpu
Artifact: assets/benchmark_metrics.json
Hardware: CPU · PyTorch 2.2.2+cpu · Windows host (platform recorded in JSON)
| Metric | BEFORE | AFTER | Δ |
|---|---|---|---|
| Smooth Manhattan wirelength | 15,540.73 | 14,170.22 | −8.82% |
| Total overlap area | 6,410.40 | 55.16 | −99.14% |
| Overlap pairs | 31 | 12 | −19 |
| Cells with ≥1 overlap | 26 / 32 | 17 / 32 | — |
| Runtime | — | 2.874 s | — |
Side-by-side: 
Residual overlaps after 2000 epochs are expected on this tiny, densely interconnected toy — the optimizer prioritizes clearing bulk overlap area while keeping nets short.
git clone https://github.com/ArchanaChetan07/GradPlace.git
cd GradPlace
python -m venv .venv
# Windows: .\.venv\Scripts\Activate.ps1
source .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt
python scripts/eval_before_after.py --epochs 2000 --outdir assets --device cpuDocker:
docker compose up --build
# plots + assets/benchmark_metrics.json refreshed in ./assetsCUDA optional: --device cuda when a GPU is available; CI and Docker default to CPU.
pytest tests/ -v --tb=shortVerified this session: 6/6 passed — bin-cap constant, λ ramp schedule, monotonicity, CPU placer smoke on the synthetic toy, and metrics JSON schema.
| Layer | Technology |
|---|---|
| Placer | PyTorch (placement.py) |
| Viz | matplotlib |
| Eval | scripts/eval_before_after.py |
| Containers | Dockerfile · Compose |
| CI | GitHub Actions (CPU torch + pytest + Docker build) |
See repository.
Archana Chetan · @ArchanaChetan07